Date:	Thu, 18 Mar 2010 14:04:32 -0700
From:	"Paul E. McKenney" <paulmck@...ux.vnet.ibm.com>
To:	Mathieu Desnoyers <mathieu.desnoyers@...icios.com>
Cc:	linux-kernel@...r.kernel.org, mingo@...e.hu, laijs@...fujitsu.com,
	dipankar@...ibm.com, akpm@...ux-foundation.org,
	josh@...htriplett.org, dvhltc@...ibm.com, niv@...ibm.com,
	tglx@...utronix.de, peterz@...radead.org, rostedt@...dmis.org,
	Valdis.Kletnieks@...edu, dhowells@...hat.com,
	eric.dumazet@...il.com, Alexey Dobriyan <adobriyan@...il.com>,
	Peter Zijlstra <a.p.zijlstra@...llo.nl>
Subject: Re: [PATCH tip/core/urgent 2/2] rcu: remove INIT_RCU_HEAD,
 RCU_HEAD_INIT, RCU_HEAD

On Thu, Mar 18, 2010 at 04:22:02PM -0400, Mathieu Desnoyers wrote:
> * Paul E. McKenney (paulmck@...ux.vnet.ibm.com) wrote:
> > On Thu, Mar 18, 2010 at 03:35:20PM -0400, Mathieu Desnoyers wrote:
> > > * Paul E. McKenney (paulmck@...ux.vnet.ibm.com) wrote:
> > > > From: Alexey Dobriyan <adobriyan@...il.com>
> > > > 
> > > > call_rcu() will unconditionally reinitialize RCU head anyway.  New users
> > > > of these macros constantly appear, so remove them.
> > > 
> > > Hrm. So do we have something that checks for double-use of a RCU head at
> > > the moment ? (using call_rcu() twice on the same head without being
> > > certain that the first callback have finished its execution).
> > > 
> > > I think that hiding rcu head initialization into call_rcu() is one more
> > > step towards misuses that will silently corrupt rcu head lists. So I
> > > think we should first add the double-use debugging option before we
> > > remove the RCU head initializations.
> > 
> > So your thought is to have rcu_do_batch() do something like the
> > following?
> > 
> > 	...
> > 
> > 	next = list->next;
> > 	prefetch(next);
> > 	list->next = RCU_HEAD_INIT_PTR;
> > 	func = list->func;
> > 	list->func = RCU_HEAD_INIT_PTR;
> > 	func(list);
> > 	... /* touching anything referenced by "list" is use-after-free. */
> > 
> > Then have __call_rcu() do something like the following before initializing
> > the ->func and ->next pointers:
> > 
> > 	WARN_ON_ONCE(head->next != RCU_HEAD_INIT_PTR ||
> > 		     head->func != RCU_HEAD_INIT_PTR);
> > 
> > And then require that all users of call_rcu() and friends use one of the
> > RCU_INIT() macros?
> > 
> > Or did you have something else in mind?
> 
> More precisely poisoning an extra field of the rcu_head, as done in the
> following patch.
> 
> I posted it a few months ago, but has been rejected on the ground that
> it should be re-done in within the debug objects infrastructure.  But I
> had to focus on other things and never found time to do these changes.
> It needs a separate patch which adds missing INIT_RCU_HEAD() to a few
> more kernel sites.

Indeed!

> The reason why I add a supplementary field for the poison is to be able
> to warn for detection of incoherent list_head both in call_rcu and in
> rcu_do_batch(), which does not seem possible with the scheme you propose
> above. The sequence is:
> 
> init ->         debug = NULL
> call_rcu ->     WARN_ON_ONCE(debug != NULL)
>                 debug = LIST_POISON1
> rcu_do_batch -> WARN_ON_ONCE(debug != LIST_POISON1)
>                 debug = NULL
> 
> 
> tree rcu: Add debug RCU head option
> 
> Poisoning the rcu_head callback list. Only for rcu tree for now.
> 
> Helps finding racy users of call_rcu(), which results in hangs because list
> entries are overwritten and/or skipped.

This does look attractive!

Some comments below.

> Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@...icios.com>
> CC: "Paul E. McKenney" <paulmck@...ux.vnet.ibm.com>
> CC: mingo@...e.hu
> CC: akpm@...ux-foundation.org
> ---
>  include/linux/rcupdate.h |   11 +++++++++++
>  include/net/dst.h        |    2 ++
>  kernel/rcutree.c         |   10 ++++++++++
>  lib/Kconfig.debug        |    9 +++++++++
>  4 files changed, 32 insertions(+)
> 
> Index: linux-2.6-lttng/include/linux/rcupdate.h
> ===================================================================
> --- linux-2.6-lttng.orig/include/linux/rcupdate.h	2009-11-22 20:25:49.000000000 -0500
> +++ linux-2.6-lttng/include/linux/rcupdate.h	2009-11-22 22:11:48.000000000 -0500
> @@ -49,6 +49,9 @@
>  struct rcu_head {
>  	struct rcu_head *next;
>  	void (*func)(struct rcu_head *head);
> +#ifdef CONFIG_DEBUG_RCU_HEAD
> +	struct rcu_head *debug;
> +#endif
>  };
> 
>  /* Exported common interfaces */
> @@ -77,11 +80,19 @@ extern int rcu_scheduler_active;
>  #error "Unknown RCU implementation specified to kernel configuration"
>  #endif
> 
> +#ifdef CONFIG_DEBUG_RCU_HEAD
> +#define RCU_HEAD_INIT 	{ .next = NULL, .func = NULL, .debug = NULL }
> +#define RCU_HEAD(head) struct rcu_head head = RCU_HEAD_INIT
> +#define INIT_RCU_HEAD(ptr) do { \
> +       (ptr)->next = NULL; (ptr)->func = NULL; (ptr)->debug = NULL; \
> +} while (0)
> +#else
>  #define RCU_HEAD_INIT	{ .next = NULL, .func = NULL }
>  #define RCU_HEAD(head) struct rcu_head head = RCU_HEAD_INIT

RCU_HEAD() is identical in either case, so should be pulled out of the
#ifdef, right?

>  #define INIT_RCU_HEAD(ptr) do { \
>         (ptr)->next = NULL; (ptr)->func = NULL; \
>  } while (0)
> +#endif
> 
>  #ifdef CONFIG_DEBUG_LOCK_ALLOC
>  extern struct lockdep_map rcu_lock_map;
> Index: linux-2.6-lttng/kernel/rcutree.c
> ===================================================================
> --- linux-2.6-lttng.orig/kernel/rcutree.c	2009-11-22 21:38:56.000000000 -0500
> +++ linux-2.6-lttng/kernel/rcutree.c	2009-11-22 22:10:49.000000000 -0500
> @@ -39,6 +39,7 @@
>  #include <asm/atomic.h>
>  #include <linux/bitops.h>
>  #include <linux/module.h>
> +#include <linux/poison.h>
>  #include <linux/completion.h>
>  #include <linux/moduleparam.h>
>  #include <linux/percpu.h>
> @@ -1010,6 +1011,10 @@ static void rcu_do_batch(struct rcu_stat
>  		next = list->next;
>  		prefetch(next);
>  		trace_rcu_tree_callback(list);
> +#ifdef DEBUG_RCU_HEAD

This needs to be CONFIG_DEBUG_RCU_HEAD, right?

> +		WARN_ON_ONCE(list->debug != LIST_POISON1);
> +		list->debug = NULL;
> +#endif
>  		list->func(list);
>  		list = next;
>  		if (++count >= rdp->blimit)
> @@ -1291,6 +1296,11 @@ __call_rcu(struct rcu_head *head, void (
>  	unsigned long flags;
>  	struct rcu_data *rdp;
> 
> +#ifdef DEBUG_RCU_HEAD

Ditto here.

> +	WARN_ON_ONCE(head->debug);
> +	head->debug = LIST_POISON1;
> +#endif
> +
>  	head->func = func;
>  	head->next = NULL;
> 
> Index: linux-2.6-lttng/lib/Kconfig.debug
> ===================================================================
> --- linux-2.6-lttng.orig/lib/Kconfig.debug	2009-11-22 22:01:03.000000000 -0500
> +++ linux-2.6-lttng/lib/Kconfig.debug	2009-11-22 22:10:49.000000000 -0500
> @@ -652,6 +652,15 @@ config DEBUG_LIST
> 
>  	  If unsure, say N.
> 
> +config DEBUG_RCU_HEAD
> +	bool "Debug RCU callbacks"
> +	depends on DEBUG_KERNEL
> +	depends on TREE_RCU
> +	help
> +	  Enable this to turn on debugging of RCU list heads (call_rcu() usage).
> +	  Seems to find problems more quickly with stress-tests in single-cpu
> +	  mode.
> +
>  config DEBUG_SG
>  	bool "Debug SG table operations"
>  	depends on DEBUG_KERNEL
> Index: linux-2.6-lttng/include/net/dst.h
> ===================================================================
> --- linux-2.6-lttng.orig/include/net/dst.h	2009-11-22 20:25:49.000000000 -0500
> +++ linux-2.6-lttng/include/net/dst.h	2009-11-22 22:10:49.000000000 -0500
> @@ -154,7 +154,9 @@ static inline void dst_hold(struct dst_e
>  	 * If your kernel compilation stops here, please check
>  	 * __pad_to_align_refcnt declaration in struct dst_entry
>  	 */
> +#ifndef CONFIG_DEBUG_RCU_HEAD
>  	BUILD_BUG_ON(offsetof(struct dst_entry, __refcnt) & 63);
> +#endif

You lost me on this one.

>  	atomic_inc(&dst->__refcnt);
>  }

Would you be willing to add this to TINY_RCU as well?  It would be under
#ifdef, so would not affect the size of production builds.

							Thanx, Paul
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists